AI was supposed to make work cheaper, faster, and dramatically more productive. Instead, a new problem is emerging: people may be using more AI without creating enough additional value to justify the cost.

That is the central concern raised by Chamath Palihapitiya in his recent discussion about the economics of artificial intelligence.

The issue is not whether AI works. It clearly does. The issue is whether companies are using increasingly expensive AI in places where the productivity gains are large enough to offset the underlying costs.

That distinction matters.

For someone already juggling meetings, software subscriptions, automated tools, AI assistants, dashboards, documents, and an endless stream of new technology, the temptation is obvious: if AI can help, use more of it.

But more usage does not automatically mean more productivity.

For decades, the most powerful intelligence lived behind the closed doors of quant firms — billion-dollar funds whose algorithms quietly out-traded everyone else.

That era just ended.

Co-Invest by Liquid is the first way to trade directly inside Claude and ChatGPT. Ask your AI to analyze a market, stress-test an idea, or build a position sized to your comfort level, then execute, right there in the conversation. No jargon. No twelve-screen terminal. No guesswork.

It's built for people who want to invest smarter, not gamble harder. You set the risk tolerance. The AI does the heavy lifting. You approve every trade.

The institutions made the game, Co-Invest gives you a way to beat them.

Trade with Co-Invest Today

In fact, the next phase of the AI story may be less about convincing people to use AI and more about teaching organizations when AI is actually worth using.

Chamath’s argument starts with something that is easy to overlook when AI is discussed primarily as a productivity tool: every AI interaction has an economic cost.

Modern AI systems are generally priced around usage, with tokens serving as a measurement of the text processed by the model. The more information an organization sends into AI systems and the more output it requests, the greater the potential computing expense.

That creates an unusual situation.

A company can encourage employees to use AI because management believes it will make them more productive. Employees then start using AI everywhere: drafting documents, analyzing information, generating code, summarizing meetings, researching questions, reviewing material, creating presentations, and experimenting with increasingly sophisticated agents.

Individually, each interaction can look inexpensive.

Collectively, however, thousands of employees repeatedly using powerful models can create a substantial operating expense.

Chamath calls this behavior “tokenmaxxing”: essentially pushing AI usage aggressively because the organization believes more AI consumption should translate into more productivity.

The danger is that the connection between those two things is not automatic.

A company could dramatically increase the number of tokens it consumes while generating only a modest improvement in revenue, output, or efficiency.

That is where the productivity gap begins.

Tip: Treat AI usage as a business input, not a productivity metric. The important question is not how much AI is being used, but what measurable improvement that usage produces.

Imagine a team that previously spent 100 hours producing a particular type of work.

AI is introduced, and the team begins using it heavily. Their AI consumption rises by 500%, but the amount of human time required falls from 100 hours to only 85.

That is an improvement.

But it may not be a very good economic improvement if the additional AI costs are significant and the resulting work does not generate enough additional value.

This is the distinction Chamath is emphasizing.

The technology can become dramatically more capable while the financial case for using it everywhere remains unclear.

That may sound counterintuitive because AI demonstrations often focus on what the technology can accomplish. A model can write software, summarize a 200-page document, generate a presentation, analyze a spreadsheet, or research a complicated question in seconds.

But the real organizational question is different:

How much economic value does that capability create after accounting for the cost of using it?

That calculation becomes particularly important as companies move from occasional AI assistance toward continuous AI usage.

If an employee uses AI once to eliminate two hours of tedious work, the economics are relatively easy to understand.

If an organization builds an AI-heavy workflow that continuously consumes large quantities of tokens, invokes multiple models, retries failed outputs, processes large context windows, stores information, and uses agents to perform additional tasks, the cost structure becomes much harder to see.

The technology may feel inexpensive at the individual-user level while becoming significant at the organizational level.

That is exactly why Chamath believes some companies could eventually discover unexpected increases in operating expenses.

One of the most interesting parts of Chamath’s argument is that AI spending may initially be difficult for executives to see clearly.

Traditional software spending is comparatively straightforward. A company buys a certain number of licenses, pays a known subscription fee, and can generally understand what it is spending.

AI can behave differently.

Usage can vary enormously between employees, teams, applications, models, and workflows. A single application can generate dramatically different costs depending on how much context it processes, how often it calls a model, which model it uses, and how many times an agent repeats a task.

That creates the possibility of distributed AI spending.

One employee might use AI casually. Another might rely on it throughout the day. A development team might run thousands of automated requests. Another department might build an internal agent that continuously processes information.

None of those decisions necessarily looks alarming in isolation.

Together, they can become a meaningful operating expense.

Chamath's warning is essentially that executives could eventually discover the cost through financial results rather than through a neatly presented technology budget.

A company misses its earnings expectations by a few cents. Management investigates the unexpected expense. Eventually, someone discovers that AI usage has expanded much faster than anticipated.

That scenario is not a prediction that every company will experience an AI cost crisis. It is a warning about what happens when usage grows faster than financial controls.

Tip: Make AI spending visible at the workflow level. Knowing which teams use AI is useful; knowing which workflows create measurable value is far more useful.

Chamath also uses an interesting analogy: AI is increasingly becoming a commodity.

His comparison is essentially to oil.

Different producers may offer different grades and prices, but the basic idea is that intelligence is becoming available at increasingly low costs.

That creates a major problem for companies building workflows around premium models.

Suppose a highly capable model costs substantially more than a competing model, but the cheaper model performs nearly as well for the task being completed.

Why would a company continue paying the premium?

The answer should be simple: because the additional capability produces additional economic value.

If it does not, the premium becomes difficult to justify.

This is why model convergence matters.

Chamath argues that the gap between leading AI models is narrowing for many everyday applications. A model does not necessarily need to be the absolute best available to be useful. If a lower-cost system can perform 80%, 90%, or 95% as well for a particular task, companies have a strong economic reason to consider it.

The most expensive model may still be the right choice for highly specialized work where accuracy or reasoning quality directly affects revenue or risk.

But using the most powerful model for everything is increasingly difficult to defend.

It is similar to choosing a professional-grade tool for every task simply because it is technically better. The better tool is not automatically the better economic decision.

Tip: Match model quality to task value. Reserve expensive, high-capability models for work where their additional performance actually changes the outcome.

This is where AI economics gets particularly interesting.

For years, technological competition was largely about building the best model.

But if models continue to converge in practical performance, the competition can shift toward cost, availability, reliability, speed, and scale.

Imagine two models.

Model A is slightly better but costs significantly more to operate.

Model B is slightly less capable but handles 95% of the same tasks at a fraction of the cost.

For many businesses, Model B may be the economically superior choice.

That does not mean frontier models become irrelevant. There will always be tasks where the best available reasoning, coding, research, or specialized capabilities justify higher costs.

But most organizations do not spend every minute solving the hardest possible problem.

Much of everyday work consists of routine analysis, drafting, classification, summarization, transformation, search, and repetitive processes.

Those tasks may not require the most expensive intelligence available.

This creates an emerging principle for AI adoption:

Use the cheapest capable system, not the most impressive system.

That principle could become just as important as model accuracy.

There is another trap hiding underneath the AI boom.

AI makes it incredibly easy to produce more. More emails. More documents. More code. More presentations. More analyses. More ideas. More reports. More automated workflows.

But producing more output does not necessarily mean producing more value.

A person can generate a 30-page report in minutes and still have created something nobody needed.

A development team can generate thousands of lines of code faster than ever and still increase maintenance problems.

A company can automate dozens of workflows and still spend more time reviewing, correcting, and coordinating the resulting output. This is why the AI productivity debate cannot simply be measured by speed.

The more useful question is: What became better because AI was introduced?

Did customers receive a better experience? Did employees eliminate genuinely unnecessary work? Did the company increase revenue? Did operating costs decline? Did quality improve? Did important decisions become faster or more accurate? Did teams gain capacity to work on higher-value problems? Those are productivity outcomes. The number of prompts sent to an AI model is not.

Tip: Measure the finished outcome rather than the amount of AI activity required to produce it.

partnerUP removes the bottleneck. You guide the brief, it does the rest. Outreach at a scale no one person could reach, scoring every applicant by fit, then handling contracts, payments, and tracking in one place. Your first two creators are on us.

Start with two free

The phrase itself captures an important behavioral problem.

Once AI becomes available, employees may naturally begin finding more ways to use it.

That sounds harmless.

But consider what happens when AI becomes embedded into every step of a workflow.

Instead of thinking through a problem and then asking AI for assistance, someone may ask AI first.

Instead of writing a rough draft and using AI to improve it, they may ask AI to generate the entire document.

Instead of debugging a problem and using AI as a second opinion, they may immediately hand the problem to an agent.

The organization becomes increasingly dependent on AI consumption.

At that point, AI is no longer simply helping people work.

It is becoming part of the fundamental operating process.

That makes cost discipline more important, because eliminating the AI expense may also mean redesigning the workflow itself.

And there is a second issue: dependency can hide inefficiency.

If AI makes a bad process faster, the organization has not necessarily become more productive. It may simply be producing the same inefficiency at greater speed.

The early phase of AI adoption was dominated by a simple question:

“Are you using AI?”

That question is becoming outdated.

The more useful question is:

“Where does AI create enough value to justify its cost?”

That requires a different mindset.

Instead of rolling out AI across an organization and encouraging everyone to maximize usage, companies can identify specific workflows where AI has a measurable advantage.

Customer support is one example.

If AI can resolve simple customer questions automatically while allowing human employees to focus on complicated cases, the productivity gain can be measured.

Software development offers another.

If AI reduces the time required to implement routine features without increasing defects or maintenance costs, the benefit becomes measurable.

Research and analysis can be another.

If AI allows a team to examine more information without reducing the quality of its conclusions, the additional capacity can have genuine value.

The strongest use cases are therefore not necessarily the flashiest ones.

They are the ones where the economics are obvious.

Chamath is not arguing that expensive models should disappear.

In fact, his argument makes the opposite point.

If an expensive model produces enough additional value, use it.

Consider a cybersecurity company protecting critical infrastructure. If a more capable AI system materially improves threat detection and helps generate billions of dollars in additional revenue or prevents enormous losses, the higher AI cost may be trivial relative to the value created.

The same logic can apply to highly specialized research, advanced engineering, complex software development, or other situations where a marginal improvement in capability has an outsized economic impact.

The mistake is assuming that every task deserves the same level of intelligence.

You do not need a race car to drive to the grocery store.

You need the race car when the race matters.

Tip: Ask what the additional intelligence buys you. If the premium model produces a meaningfully better business outcome, the higher cost may be justified; if not, downgrade.

This may be the deeper lesson behind the entire discussion.

AI capabilities are advancing extremely quickly.

Organizational processes are not.

A company can purchase access to a powerful model almost instantly. Changing how thousands of employees work is considerably harder.

That creates a mismatch.

Technology improves in months.

Processes may take years.

Training, governance, budgeting, security, measurement, data architecture, and workflow redesign all take time.

So the bottleneck may increasingly move away from AI capability itself.

The technology may be ready.

The organization may not be.

For someone navigating this shift, that means there is little value in chasing every new AI release simply because it is more capable.

The greater advantage may come from learning how to integrate AI selectively into the work that actually matters.

The first AI race was about adoption.

The next one may be about efficiency.

Companies are likely to become more sophisticated about which models they use, how often they call them, how much context they provide, and which tasks should remain human-led.

That does not mean AI enthusiasm is ending.

It means the conversation is becoming more mature.

AI can absolutely increase productivity. But productivity is not created simply because a model is powerful or because employees use it frequently.

The value appears when the economic output created by AI exceeds the cost of deploying it.

That sounds obvious, but the speed of the AI boom makes it surprisingly easy to forget.

For you, the practical lesson is straightforward: you do not need to use AI everywhere. You need to understand where it genuinely improves the work.

The best AI workflow may not be the one with the most automation.

It may be the one that knows exactly where automation stops being worth it.

Tip: Before adding another AI tool, ask three questions: What problem does it solve? What measurable value does it create? And is that value greater than its full cost?

The biggest change ahead may not be another dramatic jump in model capability.

It may be a change in how people think about AI.

The novelty of having an AI assistant is fading. AI is becoming infrastructure. And once a technology becomes infrastructure, economics matter more than excitement.

The companies that benefit most may not be the ones that consume the most tokens, subscribe to the most tools, or deploy the most agents.

They may be the ones that understand the difference between AI activity and actual productivity.

That distinction gives you a useful filter in an increasingly noisy environment. You do not have to chase every model. You do not have to automate every task. You do not have to maximize AI usage simply because everyone else is doing it.

Instead, look for the moments where AI removes meaningful friction, increases valuable output, improves quality, or creates capabilities that were previously too expensive or time-consuming.

That is where the cost-versus-productivity equation starts working in your favor.

And as Chamath’s warning suggests, the organizations that ignore that equation may eventually discover that becoming more AI-powered did not necessarily make them more productive.

It simply made their technology bill bigger.

What’s your next spark? A new platform engineering skill? A bold pitch? A team ready to rise? Share your ideas or challenges at Tiny Big Spark. Let’s build your pyramid—together.

That’s it!

Keep innovating and stay inspired!

If you think your colleagues and friends would find this content valuable, we’d love it if you shared our newsletter with them!

As the world becomes increasingly digital, this question will be on the minds of millions of people seeking new income streams in 2026.

The answer is—Absolutely!

Thank you for taking the time to read today’s email! Your support allows me to send out this newsletter for free every day. 

 What do you think for today’s episode? Please provide your feedback in the poll below.

Share the newsletter with your friends and colleagues if you find it valuable.

Disclaimer: The "Tiny Big Spark" newsletter is for informational and educational purposes only, not a substitute for professional advice, including financial, legal, medical, or technical. We strive for accuracy but make no guarantees about the completeness or reliability of the information provided. Any reliance on this information is at your own risk. The views expressed are those of the authors and do not reflect any organization's official position. This newsletter may link to external sites we don't control; we do not endorse their content. We are not liable for any losses or damages from using this information.